Week 1 by the Numbers: A Model's Look at the Opening NFL Slate
A Bayesian score-distribution model picks seven Week 1 NFL games inside a disciplined 2–10% edge band, and flags everything outside it.
Writing
// notes on data, products, and things I've shipped
The model converged. The numbers didn't exist.
My AI assistant reported clean convergence diagnostics for a Bayesian model. The values didn't appear in any output file.
I shipped a burn mechanism. Then I built the thing that tells me if it works.
I shipped a burn mechanism. Then I built the thing that tells me if it works. A token burn is a claim, not a result.
My NFL model forecasts the line, not the game
63% closing line value across 3,331 out-of-sample bets, monotone in edge size, robust to the obvious confound — and worth nothing.
How I proved my model couldn't get better, then spent four attempts proving it again
A Murphy decomposition showed my UFC model was already perfectly calibrated and that everything left was missing information, not modeling error.